Automated Rheumatic Heart Disease Detection from Phonocardiogram in Cardiology Ward Melkamu Hunegnaw Asmare1,2,3, Frehiwot Woldehanna3, Luc Janssens1 and Bart Vanrumste1,2 1KU Leuven, Campus Group T, eMedia Research Lab, Leuven, Belgium 2KU Leuven, Electrical Engineering Department (ESAT), STADIUS, Leuven, Belgium 3Addis Ababa Institute of Technology, Center of Biomedical Engineering, Addis Ababa, Ethiopia [email protected] Keywords: Rheumatic Heart Disease, Machine Learning, Support Vector Machines, Phonocardiogram. Abstract: Rheumatic Heart Disease (RHD) is a preventable and treatable form of cardiovascular diseases. It is also re- ferred to as the ailment of the disadvantaged mainly affecting children and young adults. RHD is recognized as a global health priority by World Health Organization. This chronic heart condition silently deteriorates the normal function of the heart valves which can be detected as a heart murmur using a stethoscope. As the cardiac auscultation process is an elusive process, the clinician will always be tempted to refer the patient for expensive and sophisticated imaging procedures like echocardiography. In this study, a machine learning algorithm is developed to augment the limitation in the auscultation process and transform the stethoscope as a powerful screening tool. For this current study, an RHD heart sound data set is recorded from one hundred seventy subjects. A total of twenty-six features are extracted to model murmur due to RHD. Twenty-four classification and regression algorithms have been tested out of which the Cubic SVM has demonstrated su- periority with a classification accuracy of 97.1%, with 98% sensitivity, 95.3 % of specificity 97.6% precision. The corresponding positive predictive values (PPV) are 96% and 97% for normal and RHD respectively. The results are based on data collected from a cardiology ward where there are more pathological cases than con- trols. Hence it is a valuable detection tool in a cardiology clinic. But in the future, integrating this machine learning algorithm with a mobile phone can be a powerful screening tool in places where access to echocar- diography and cardiologist is difficult. Thus, it can then aid a timely, affordable and reliable detection tool allowing a non-medically trained individual to screen and detect RHD. 1 INTRODUCTION normally found in the skin and in the throat of healthy people. GAS is an important cause of throat infection. In certain susceptible people, usually children, the Due to the rapid epidemiological transition observed immune system becomes confused and attacks both in developing countries, not only communicable dis- the GAS bacteria and parts of the host’s body. This eases but also noncommunicable diseases are autoimmune at- tack causes the inflammation of the becoming the major cause of death risks. joints, skin, brain and most importantly the heart Cardiovascular diseases, cancer, chronic respiratory (Watkins DA, 2017). RHD is a chronic heart disorders, and diabetes are the most common ones. condition and early in the dis- ease, there are usually Among these, cardiovascular disorder takes the no symptoms. The disease can silently progress leading role (WHOAnnualReport, 2013). Worldwide, especially after repeated episodes of infection. Each ischemic heart disease is the number one cause of episode brings renewed heart valve inflammation that death, which affects males with age usually 65 or eventually leads to local scarring and distortion of the more (Emelia J. Benjamin, 2019). However, RHD is valve architecture. First, the affected valve starts to the leading cause of cardiovascular disorders in leak, normally referred to as re- gurgitation; later the middle- and low-income countries. The average age scarring can stop the valve from opening properly is around 28 years with females affected twice as and make it narrow for sufficient blood passage, much as men (Watkins DA, 2017). referred to as stenosis. These abnormalities create RHD is caused by Group A streptococcal unusually turbulent blood flow in the heart chambers Bacteria (GAS) infection. These bacteria are which is called heart murmur. Left untreated, RHD also assembles the largest heart sound dataset. will compromise the cardiac output of the patient The 2016 PhysioNet Computing in Cardiology which will subsequently lead to pre- mature death Challenge was one of the most successful challenges (Walsh, 2019). The heart sound wave- form has conducted by the program which attracted a large distinct features called the first heart sound (S1), the number of researchers to solve the heart sound classi- second heart sound (S2), systole and diastole parts. fication to normal and abnormal. In the competition, Murmur normally presents itself in the systolic or the largest heart sound dataset compiled by (Liu C1, diastolic parts. The heart sound can be listened to 2016) was provided. The winners of the competi- and recorded using a stethoscope in the form of a tion, Potes et al. (Cristhian Potes, 2016) have devel- phonocardiograph (PCG). A three second MATLAB oped a deep-learning-based classifier that combines plot of clean, noisy and murmur types of heart sound time-frequency features with a reported sensitivity of is shown in Fig. 1. The clean signal is manually 96%, specificity of 80% and overall accuracy of 89%. segmented to locate S1, S2, systole and diastole. Almost all previously proposed algorithms needed the segmentation of the heart sound recording into first heart sound, second heart sound, systole, and di- astole parts. This is a reasonable assumption which may lead to pinpointing of abnormalities in the heart sounds at specific temporal locations. However, the complexity and also the error introduced in the accu- rate localization of the segments have decreased the performance of the algorithms. Recently, P. Langley and A. Murray (Cristhian Potes, 2017) have demonstrated the feasibility of accurate classification without segmentation of the Figure 1: Time series representation of heart sounds: clean heart sounds. The paper has a relatively lower overall heart sound with S1, S2, Systole and Diastole labelled (top), accuracy of 79% (specificity 80%, sensitivity 77%) noisy heart sound (middle), and heart sound with a murmur (bottom). classification, and claims this is mainly due to the quality of the dataset used. Despite the sheer volume The heart sound gives vital information about of research done in the area, the studies are critically cardiac wellbeing. However, even under ideal condi- hampered by the lack of high- quality recordings that tions, the accuracy of diagnosis is very low (Pelech, have proper validation and standardization. This 2004). This is in reality attributed to the inherent would have created common formatting that allows limitation of the human auditory system to perform collaborative research, large- scale analytics, and accurate auscultation. On top of that, the listening tools and methodologies to be shared. The largest process is highly subjective. This usually forces available open access data set is available which was doctors to be highly dependent on other expensive compiled by Liu et al. (Liu C1, 2016). It contains 2435 imaging devices like echocardiography and x-ray for heart sound recordings from 1297 subjects. The cardiac screening (Vukanovic-Criley JM, 2006). dataset consists of recordings from subjects with a variety of abnormalities which include heart valve To counter the subjectivity and the high percent- damage and coronary artery disease. The maximum age of diagnostic errors, computer-aided diagnostic overall accuracy reported in the literature by using this (CAD) systems can provide paramount importance database is only 94% which was achieved by (Bozˇo Tomas, 2007), (Belloni and Spoletini, 2007). introducing different model optimization techniques For the successful implementation of CADs, the qual- (Suhm, 2019). ity of the input signal should be high. Such automa- D.B. Springer et al. (D.B. Springer, 2014) have tion has been researched for over six decades now. In worked on a dataset that is recorded to classify an the 1960s, one of the ground-breaking studies in the RHD from normal heart sounds. A total of 318 automatic classification of heart sound pathology was recordings from 106 subjects where 40 were identi- performed by (D. S. Gerbarg and Hofler, 1963). Since fied with RHD. Their aim was to detect systolic mur- then thousands of research papers have been mur hence the heart sound is segmented before feature published. Some of the prominent works have been extraction. A combination of MFCC and wavelet fea- properly investigated in the report published in 2016 tures are used. SVM classification algorithm is used by (Liu C1, 2016). This paper demonstrates the im- by optimizing its parameters and the procedure is portance of a well-characterized dataset for develop- validated using a 10-fold cross-validation technique. ing successful classification algorithms. This work They reported a maximum F1 score of only 0.7, the ure, the S1 and the S2 are clearly identifiable. Fig. 2 sensitivity of 74.8% and specificity of 74.5%. Poor (bottom) shows the corresponding spectrogram to vi- quality recording and external generator noise were to sualize how the energy is distributed over time. Clicks blame for such low performance. This demonstrates and glitches which are common features of a murmur the necessity for a large and reliable dataset which can be very well visualized in the spectrogram. takes